Weißenborn et al. (2025) Neural networks in catchment hydrology: a comparative study of different algorithms in an ensemble of ungauged basins in Germany

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-10-14
• Authors: Max Weißenborn, Lutz Breuer, Tobias Houska
• DOI: 10.5194/hess-29-5131-2025

Research Groups

• Institute for Landscape Ecology and Resources Management (ILR), Research Centre for BioSystems, Land Use and Nutrition (IFZ), Justus Liebig University Giessen, Germany
• Centre for International Development and Environmental Research (ZEU), Justus Liebig University Giessen, Germany
• Institute of Soil Science and Site Ecology, TU Dresden, Germany

Short Summary

This study comparatively evaluates Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) for daily discharge prediction in 35 ungauged basins in Hesse, Germany. It finds that all models show significant predictive capabilities, with CNN exhibiting slightly superior accuracy, while GRU offers the best computational efficiency, and the inclusion of static catchment features consistently improves performance.

Objective

• To evaluate the potential of predicting discharge in ungauged basins using daily forcing data with LSTM, CNN, and GRU neural networks.
• To compare the computational efficiency of LSTM, CNN, and GRU models for daily time series prediction.
• To investigate the potential of static catchment features to enhance prediction performance.
• To assess the impact of batch size on model performance and computational efficiency.

Study Configuration

• Spatial Scale: 89 catchments in Hesse, Germany (54 for training, 35 for testing).
• Temporal Scale: Daily resolution. Training data covered 28 years (1991–2018), and testing data covered 6 years (1997–2002).

Methodology and Data

• Models used: Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU).
• Data sources:
  • Dynamic forcing data: Daily sums of precipitation (mm), daily sums of evapotranspiration (mm), and soil temperature at 5 cm depth (°C).
  • Static catchment attributes (11 features): Soil type, soil texture, geology type, land use, permeability, average precipitation (annual mean, mm), catchment size (m²), elongation ratio, soil depth (m), average slope (°), and average evapotranspiration (annual mean, mm).
  • Target variable: Daily discharge (mm) from gauging stations.
  • Data derived from Jehn et al. (2021) and Jehn (2020) repository, with discharge data from the Hessian Agency for Nature Conservation, Environment and Geology.

Main Results

• All three neural network architectures demonstrated significant predictive capabilities for discharge in ungauged basins.
• The CNN model with static features and a batch size of 256 achieved the highest mean Kling–Gupta Efficiency (KGE) of 0.80. LSTM and GRU models showed slightly lower mean KGEs of 0.78 and 0.77, respectively, under similar configurations.
• The integration of static catchment features consistently improved model performance across all architectures and batch sizes.
• Smaller batch sizes (256) generally led to better model performance compared to larger batch sizes (2048).
• The GRU model (batch size 256, with static features) was the most computationally efficient, being 41% faster than the CNN model and 59% faster than the LSTM model.
• Performance differences among the models were modest, with less than 3.9% disparity in KGE.
• Evaluation across flow segments revealed that KGE generally increased from lowest to highest flows. LSTM demonstrated superior consistency and generalization for lowest flow conditions (Q1 and Q2), while CNN excelled at peak flow events (Q4).
• Sensitivity analysis indicated precipitation as the primary driver for discharge prediction across all models. LSTM and GRU models showed expected negative impacts from evapotranspiration and soil temperature, whereas the CNN model exhibited a counterintuitive small positive impact from evapotranspiration.

Contributions

• Provides a comprehensive comparative analysis of CNN, LSTM, and GRU architectures for discharge prediction in ungauged basins, addressing a significant gap in hydrological literature.
• Offers valuable insights into the relative strengths and limitations of each neural network model, guiding future applications and development in hydrological prediction.
• Conducts a detailed sensitivity analysis to identify key drivers affecting model predictions, contributing to refined model selection and calibration strategies.
• Evaluates the impact of static catchment features and batch size on model performance and computational efficiency, demonstrating their critical role in enhancing generalization capacity, especially in data-scarce ungauged basins.

Funding

Not specified in the paper.

Citation

@article{Weißenborn2025Neural,
  author = {Weißenborn, Max and Breuer, Lutz and Houska, Tobias},
  title = {Neural networks in catchment hydrology: a comparative study of different algorithms in an ensemble of ungauged basins in Germany},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-5131-2025},
  url = {https://doi.org/10.5194/hess-29-5131-2025}
}